Created by: sucuicong

训练部分主要代码如下：

def train():

    logger.info("start train YOLOv3, train params:%s", str(train_parameters))

    logger.info("create place, use gpu:" + str(train_parameters['use_gpu']))

    logger.info("build network and program")

    place = fluid.CUDAPlace(0) if train_parameters['use_gpu'] else fluid.CPUPlace()
    exe = fluid.Executor(place)

    scope = fluid.Scope()
    train_program = fluid.Program()
    start_program = fluid.Program()
    test_program = fluid.Program()
    
    feeder, reader, loss = build_program_with_feeder(train_program, start_program, place)

    pred = build_program_with_feeder(test_program, start_program, istrain=False)
    
    test_program = test_program.clone(for_test=True)
    
    train_fetch_list = [loss.name]
    
    exe.run(start_program, scope=scope)
    
    load_pretrained_params(exe, train_program)
    
    if train_parameters['print_params']:
        param_delimit_str = '-' * 20 + "All parameters in current graph" + '-' * 20
        print(param_delimit_str)
        for block in train_program.blocks:
            for param in block.all_parameters():
                print("parameter name: {}\tshape: {}".format(param.name,
                                                             param.shape))
        print('-' * len(param_delimit_str))
    
    pruned_params = train_parameters['pruned_params'].strip().split(",")
    logger.info("pruned params: {}".format(pruned_params))
    pruned_ratios = [float(n) for n in train_parameters['pruned_ratios'].strip().split(",")]
    logger.info("pruned ratios: {}".format(pruned_ratios))
    
    logger.info("build executor and init params")
    
    pruner = Pruner()
    train_program = pruner.prune(
        train_program,
        scope,
        params=pruned_params,
        ratios=pruned_ratios,
        place=place,
        only_graph=False)[0]
    
    base_flops = flops(test_program)
    test_program = pruner.prune(
        test_program,
        scope,
        params=pruned_params,
        ratios=pruned_ratios,
        place=place,
        only_graph=True)[0]
    pruned_flops = flops(test_program)

    stop_strategy = train_parameters['early_stop']
    rise_limit = stop_strategy['rise_limit']

    min_loss = stop_strategy['min_loss']
    # stop_train = False
    rise_count = 0
    total_batch_count = 0
    current_best_f1 = 0.0
    train_temp_loss = 0
    current_best_pass = 0
    current_best_box_pass = 0
    current_best_recall = 0
    current_best_precision = 0
    current_best_box_recall = 0
    current_best_box_precision = 0
    current_best_box_f1 = 0
    for pass_id in range(train_parameters["num_epochs"]):
        logger.info("current pass: {}, start read image".format(pass_id))
        batch_id = 0
        total_loss = 0.0
        for batch_id, data in enumerate(reader()):
            t1 = time.time()
            loss = exe.run(train_program, feed=feeder.feed(data), fetch_list=train_fetch_list)
            period = time.time() - t1
            loss = np.mean(np.array(loss))
            total_loss += loss
            batch_id += 1
            total_batch_count += 1
            
            if batch_id % 200 == 0:
                logger.info("pass {}, trainbatch {}, loss {} time {}".format(pass_id,
                                                                             batch_id, loss, "%2.2f sec" % period))
        pass_mean_loss = total_loss / batch_id
        logger.info("pass {0} train result, current pass mean loss: {1}".format(pass_id, pass_mean_loss))

    logger.info("end training")`

######################################################### #####################报错信息如下：

Python Call Stacks (More useful to users):
------------------------------------------
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/framework.py", line 2594, in _prepend_op
    attrs=kwargs.get("attrs", None))
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/nn.py", line 5472, in autoincreased_step_counter
    attrs={'step': float(step)})
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/learning_rate_scheduler.py", line 48, in _decay_step_counter
    counter_name='@LR_DECAY_COUNTER@', begin=begin, step=1)
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/learning_rate_scheduler.py", line 387, in piecewise_decay
    global_step = _decay_step_counter()
  File "train.py", line 265, in optimizer_momentum_setting
    learning_rate=fluid.layers.piecewise_decay(boundaries=boundaries, values=values),
  File "train.py", line 371, in get_loss
    optimizer = optimizer_momentum_setting()
  File "train.py", line 306, in build_program_with_feeder
    loss = get_loss(model, outputs, gt_box, gt_label, main_prog)
  File "train.py", line 403, in train
    feeder, reader, loss = build_program_with_feeder(train_program, start_program, place)
  File "train.py", line 544, in <module>
    train()

----------------------
Error Message Summary:
----------------------
InvalidArgumentError: The Tensor in the increment Op's Input Variable X(@LR_DECAY_COUNTER@) is not initialized.
  [Hint: Expected t->IsInitialized() == true, but received t->IsInitialized():0 != true:1.] at (/paddle/paddle/fluid/framework/operator.cc:1264)
  [operator < increment > error]